3D CNN application code improvements

[oleschwen]

Following #34, we could consider the following for the application code:

• clean up unused dataset classes, methods, files, etc.
• remove unused options for datasets, models, binary vs. ordinal etc. or make sure to run only combinations (incl. loaded data) that make sense
• decide which datasets/models etc. we want to support and keep for now
• consider generating the synthetic dataset inside the container (considering we do not want to write to /data)
• consider generic way of computing numbers of epochs per site depending on (relative to full dataset size we would need to know?). This should balance the time per epoch acros sites. For a comparison of swarm vs. local training, the number of epochs in local training should be the same as in swarm training at each site.
• …

We may want to split this issue.

[oleschwen]

• do not rely on being able to download model weights (maybe, if license permits, put in Docker?; start training from scratch?)
• MST works in one variant tested
• ResNet variants to be tested
• deal with redundancy of, e.g., configured model and model parameters (e.g., empty kwargs) in config/config_fed_client.conf and the code
• deal with redundancy in swarm config
• check when building docker? use own config file and distribute to current files? rely on catching problems in manual testing?
• how to abort when test has failed and continues printing message

[oleschwen]

• output statistics of label distribution in test and validation data

[oleschwen]

ResNet 10, 18, 34, 50 work; 101 and 152 remain at AUROC 0.5.

[oleschwen]

see #16 regarding refactoring application code

[SteffenRenisch]

evaluation of final model on test data:

• ideally, we would like to habe the abilitites to
• run the evaluation automatically after the training finished (maybe even with the set of evaluation clients being different from the set of training clients - for leave-one-out type of evaluation), but also
• run the evaluation as a separate workflow